Enhanced Financial Time-Series Forecasting with Hybrid PSO-Fuzzy Rough Set-GRU Model

Abstract

Financial markets play an important role in the economic and social structure of contemporary society, with information being a valuable commodity. Yet the sheer volume of data available presents difficulties in analyzing financial assets. In this paper, we examine the effect of different deep learning models—GRU, LSTM, and CNN-LSTM- as well as a hybrid Particle Swarm Optimization (PSO) model integrated with Fuzzy Rough Set theory (FRS) and GRU in stock price prediction. From historical prices of stocks, we preprocess data and normalize feature vectors using MinMax scaling. For each model, extensive training and testing is conducted, using the performance as judged by MAE (Mean Absolute Error), RMSE (Root Mean Squared Error) and the R² (coefficient of determination). Our results indicate that the PSO-Fuzzy Rough Set-GRU model performs greatly better than the others, having the lowest Training MAE (1.56), Testing MAE (2.10), Training RMSE (2.00), and Testing RMSE (2.82), along with the highest Training R² (0.9995) and Testing R² (0.9990). By comparison, single models such as LSTM and CNN-LSTM have higher error rates and lower R² values.